The Claim
Last week, Anthropic said the quiet part loud: artificial general intelligence isn't a decade away. It might already be here — or close enough that the distinction no longer matters.
The evidence isn't theoretical. It's sitting in their production codebase.
What Anthropic Actually Showed
Several data points from Anthropic landed in rapid succession, and taken together they paint a picture that's hard to dismiss.
80% of production code is now authored by Claude
As of May 2026, more than 80% of the code merged into Anthropic's own production codebase was written by Claude — up from low single digits when Claude Code launched just 18 months ago. The average Anthropic engineer is now merging 8x as much code per quarter as they did between 2021 and 2025.
This isn't a demo. It's their live codebase, shipped to production, maintained by an AI.
Claude Mythos found a 17-year-old zero-day
Anthropic built Claude Mythos — their most powerful model yet, and one they chose not to release publicly — and immediately discovered something alarming: it can autonomously find and exploit software vulnerabilities at a level that surpasses all but the most elite human security researchers.
In testing, Mythos independently identified a 17-year-old remote code execution vulnerability in FreeBSD that had evaded researchers for nearly two decades. It then exploited it to gain root access. Unprompted. Anthropic says they did not train the model to do this. The capability emerged anyway.
More broadly, Mythos has now found thousands of high-severity vulnerabilities across every major operating system and web browser.
A 52x training speed improvement
Anthropic runs a recurring internal benchmark: ask each new model to optimize training code to run faster. With Claude Opus 4 in May 2025, the result was roughly 3x the original speed. With Mythos Preview in April 2026, that jumped to 52x.
That's not incremental progress. That's a different class of capability.
Claude is helping build Claude
In April 2026, Anthropic published a demonstration of Claude running an open-ended AI safety research project end-to-end, using nine parallel agents. Over 800 cumulative hours, those agents recovered 97% of the performance gap on the task — autonomously. AI is now actively accelerating AI development.
Is This AGI?
Depends who you ask — and which definition you use. Anthropic is careful with the word. They don't claim Mythos *is* AGI. But they're no longer comfortable claiming it isn't.
What's harder to dispute is that something has qualitatively shifted. Anthropic co-founder Jack Clark put a number on it: he now estimates a 60% probability that recursive self-improvement — the point at which AI designs and trains its own successor with minimal human input — will occur before the end of 2028.
That's two and a half years away.
Project Glasswing: The Defensive Bet
Anthropic's response to Mythos' capabilities isn't to release it. It's to build a moat.
Project Glasswing is their answer: a defensive initiative backed by $100M in credits, partnering with AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, and the Linux Foundation to use Mythos to *harden* the world's most critical codebases before offensive capabilities proliferate elsewhere.
The move is both remarkable and sobering. Remarkable because it's a serious attempt to get ahead of the threat. Sobering because it acknowledges, implicitly, that the threat is real and near.
What This Means for Developers
If you write software for a living, this is the most important context shift since the internet. The question isn't whether AI will change how code is written — that's already happened. The question is how fast the transition will be and whether we can steer it.
A few things worth paying attention to:
Security assumptions need a reset. If Mythos can autonomously find zero-days in major operating systems and browsers, every security model built on "this would be too hard to discover" is now questionable. The attack surface just got wider for everyone — defenders and attackers alike.
Developer leverage is real, but the bar is rising. An AI that helps you merge 8x more code per quarter is extraordinary. But as AI capability accelerates, value shifts toward developers who can set direction, evaluate outputs, and reason about systems at a higher level of abstraction — not just those who ship features fastest.
Open source faces new pressure. Mythos found thousands of high-severity vulnerabilities in open source projects. The maintainers of Linux, FreeBSD, and major browsers weren't designing their security models with AI-assisted exploitation at this scale in mind. Project Glasswing is one response, but the attack-vs-defense dynamic is genuinely uncertain.
The recursive loop is the variable to watch. Anthropic explicitly framed the 80% code statistic not as a productivity win, but as evidence of a feedback loop: Claude improves, Claude writes more of the next Claude, which improves faster. That compounding is the signal. If Jack Clark's 60% estimate is even roughly right, the next few years look very different from the last few.
Where We Stand
At Kodo, we believe tools like ours exist in service of developers — to reduce friction, amplify capability, and keep humans meaningfully in the loop. The events of the past few weeks don't change that mission. If anything, they sharpen it.
The developers who will thrive in this environment aren't the ones who ignore these signals. They're the ones who engage with them — who understand what AI can and can't be trusted to do, who build systems with that in mind, and who stay close enough to the tools that they see the shifts coming.
AGI or not, something has changed. The best thing we can all do is pay attention.
What do you make of Anthropic's announcements? We're talking through the implications on [Discord](#) — come join us.